DERBi PIE

DERBi PIE

DERBi PIE: Query Searches

DERBi PIE: Query Searches

DERBi PIE: What could this do for us?

• In DERBi PIE, having an integrated semantic system could provide automated answers to questions such as:

  • Are certain sound sequences associated with certain meanings or semantic spheres?

    • Sound symbolism, etc.

  • Morphological classes or derivations?

  • How have meanings changed over time into the various branches and daughter languages?

DERBi PIE: the Problem for Today

• How to create a system of semantic properties and relationships that translates across IE languages?

Word Embeddings: Overview

• Distributional Analysis: “You shall know a word by the company it keeps.” - John Rupert Firth

• For example, you are very likely to see the words “dog” and “leash” appear near each other in a text.

• However, you are less likely to see the words “dog” and “physics” as near each other.

  • Why is that?

Word Embeddings: Overview

• There exist programming tools that allow us to generate vectors (essentially coordinates) based on a word’s position within a text and proximity to other words.

• All of the vectors that can be generated from a text would lie in a semantic space or hyperspace.

Word Embeddings: Overview

• This approach in line with what is done in present-day NLP – identifying semantic relationships through word embeddings.

• Before we can run a text through one of these tools, we must first discuss tokenization and lemmatization.

Word Embeddings: Processing the Text for Tokens

• Tokenization is the process of breaking a text into chunks.

Word Embeddings: Processing the Text for Lemmas

• Lemmatization is the process of reducing multiple forms to a single one.

Word Embeddings: Identifying the Context

• The word embeddings tool can analyze what words appear commonly near another word.

Word Embeddings: Constructing a Hyperspace

• You can tinker with how exactly a word embeddings tool generates these vectors, for example:

  • How many dimensions will it generate for each vector?

  • How far to each side of a word will it look?

  • How many times will it run through the text?

Word Embeddings: Example of a Hyperspace

Plots like these can be created with generated vectors!

Word Embeddings: Example of a Vector

• Another Latin example:

Word Embeddings: Numerical Value of Semantic Similarity

• To get a numerical representation of the similarity between two vectors, we use their cosine similarity (cosine of the angle between the vectors).
• Similarity scores range from -1 to 1, with -1 indicating opposite vectors and 1 indicating proportional vectors.

Fancy formula from Wikipedia

Word Embeddings: Example of Cosine Similarity

• This score suggests that these two words are more similar than not.

Word Embeddings: Reconstructing a Hyperspace in Languages without a Corpus

• We can construct a hyperspace for ancient languages;
• But how does one do this for languages without any known corpus, such as PIE?
• Well, how do we identify other properties of PIE?

Word Embeddings : Basic Idea

• We take two, similar properties in two (or more) related languages, which allow us to approximate an earlier state in a source language

Word Embeddings : Basic Idea

• It is in this way that we propose that the use of word embedding models created by descendant languages, to approximate an earlier state in the source language

Word Embeddings: Methods

• As you can imagine, this stuff is complicated, which is why we won’t go into much detail about the specific methods;
  • see Github for four-page paper, code, data, etc.

Word Embeddings, Problem #1: Verification

• So we take descendant hyperspaces, and reconstruct an earlier, source hyperspace based on the hyperspaces provided
• But how can we trust this methodology?
  • Obviously can’t verify the hyperspace of PIE through analysis of PIE texts!

Word Embeddings, Problem #1: Verification

Word Embeddings, Problem #2: Vectors Across Models

• Vectors generated for hyperspaces take on arbitrary values when training models
  • So ‘dog’ could = (0, 0) or (-6, 100) – these values change every time you run the model
• For this reason, we must align models (Dev et al., 2021):
  • Identify substructures across language models that remain fixed
  • Use pre-aligned word embedding models (following Joulin et al., 2018)

Word Embeddings: Methods

• Models trained on French & Spanish Wikipedia articles
• Generate aligned word vectors using fastText (Bojanowski et al., 2017)
• Filter out certain words:
  • If there is no corresponding word in the other languages
  • Non-vocabulary, including words with non-language characters
• Remaining words lemmatized, further reducing vocabulary by roughly 10%

Word Embeddings: Methods

• To relate words together and find common words:
  • Words are translated into each other’s respective languages, using Google Translate, with a Python translation library deep-translator
  • The same is done with Latin, using the Latin corpus (from CLTK [the Tesserae Project]), which is lemmatized
• Any word that cannot be lemmatized in Latin (such as Greek words) is removed from the corpus

Word Embeddings: Calculating *Latin

Word Embeddings: Calculating *Latin

Word Embeddings: the “Normal” Model

• To identify the effectiveness of the method, recall that we want to compare the *Latin with the Latin
• But how do we know how well each model performs?

Word Embeddings: How to Assess the Goodness of Your Model

• We follow Stringham & Izbicki 2020 in using the OddOneOut task, which has been demonstrated to be more accurate with languages with small corpora (LRLs), especially ones that aren’t modern
• OddOneOut task demonstrated to work for languages with corpus as small as 1800 tokens (Old Gujarati)

Word Embeddings: OddOneOut (Stringham & Izbicki 2020)

• The OddOneOut technique “measures a model’s ability to identify words that are unrelated to [a specific] category”

Word Embeddings: OddOneOut (Stringham & Izbicki 2020)

• Categories are automatically drawn from Wikidata and implemented when applicable.
• Example categories include (ibid., 181):
  • Human Biblical Figures: Abednego, Abraham, Azor, Chedorlaomer, Christ, Goliath, Hilkiah, Lo-Ammi, Matthew the Apostle, Peter, Sheba, Uz, Yael, Zerubbabel
  • Hinduism: Adharma, Ashvadamedha, Brahmin, Bhagavan, Guru, Hindu Prayer Beads, Karma, Mahātmā, Mudra, Nirvana

Word Embeddings: Results

• Our results indicate mixed success: we can create a hyperspace from the descendant languages that performs reasonably well on Latin tests with OddOneOut
• Preliminary tests show that the smaller the corpus, the better the model works, as compared to the normal one

Word Embeddings: Results

• This is not exactly a bad thing, as this is the case for many languages in our discipline!
  • If a language has a large corpus (such as Ancient Greek), we can reduce it as we’ve done here for Latin
  • To train such models, reductions of tokens are done randomly, both with regards to the number of tokens and specific tokens used

Word Embeddings: Results

• Unclear what is the “magic number” for the size of corpus to arrive at most accurate representation (compared to the normal model)
• Unclear exactly why performance of descendant models tends to decrease as the corpus size increases
  • Anton has ideas, all model-/programming-specific – see the draft for further

Recap: Descendant Model

• In this talk, we implement the Comparative Method to reconstruct the hyperspaces of languages without corpora
• Upsides:
  1. fully automated
  2. requires low CPU processing
• Downsides:
  1. doesn’t distinguish multiple senses (polysemy)
  2. only performs well on smaller corpora (compared to normal model)

Word Embeddings: Problems with Current Model (Anonymous NLP reviewers)

1. Use of Google translate suboptimal, may result in translation errors; should use either bilingual dictionaries or LLMs (think GPT) for more accurate translations
2. LLMs > Word Embeddings
   1. Vectors: polysemy (e.g., ‘bank’)
   2. Vectors: context
   3. Vectors: precision (300 vs. 175B parameters)

Future Directions: using LLM models

• We believe that our reconstruction model shows promise

• We plan to stick with the Descendant Model strategy, but:

  • Utilize LLMs (such as GPT) for modelling (hyperspace) for greater precision and differentiation of polysemy

  • Instead of Google Translate, use bilingual dictionary or LLM

    • We’ve had great success with LLMs in the parsing of data for DERBi PIE

Future Directions: using LLM models

• Add additional Romance languages;

• When happy with results, move on to other subbranches (likely Slavic or Indic) for testing;

• We believe that a good, workable model should be able to generate a *proto-vector through the centroid of the descendant language-word centers (minus those additional steps).

Future Directions: Integration into DERBi PIE

• Let’s finish with where we began – DERBi PIE

• With aligned hyperspaces generated for each descendant and reconstructed language, our models would give us coordinates for each lexeme in a language

Future Directions: Integration into DERBi PIE

• Possible question: “How semantically similar are PIE *sC- roots – as compared to related *C- roots?” (*(s)peḱ- ‘look at’)

  • Calculate cosine similarities!

• Example from English:

  • bl- words: “blue”, “blaze”, “bland”, “blush”, “blink”, “blow”, “blast”, “blot”, “blend”, “bleak”

  • These words have a cosine similarity score of .8016 (remember: perfect score is 1.0!)

Thank you!

Download: Slides, Paper, Code